To understand why they syck at real world tasks, you gotta realize what LLMs are as a model. What do they model? They model static sequences, an example of whic

To understand why they syck at real world tasks, you gotta realize what LLMs are as a model. What do they model? They model static sequences, an example of which is language, the first thing they were used on. What are those? 'Sequences' is self explanatory. 'Static' is interesting, it means that the future state of the sequence is not dependent on your current action. As opposed to dynamical, which means the next state is a function of your current action/prediction, and your next action is a function of that resulting state and so on and so on. An example of a dynamical system is easy, the stock market. If you sell 100 dollars of nvidia stock right now, it would have no effect on the market, but if you sold 50 billion dollars of nvidia stock, the market will definitely react, at least a little. You current action can change the future state of the system. An example of a static system the way LLMs are trained on sentence pairs where you could have the sentence pair "i went to the bank. Swimming in the river was alot of fun" ans you give the LLM the first sentence and ask it to tell you what type of bank the sentect is referring to. If the LLM incorrectly says financial bank, the second sentence doesnt magically change to talk about a cashing a check, it stays the exact same. It stays the same whther the LLM guesses right or wrong. You just tell it whether its right or wrong If your not scared of equations, a static system is y = f(x), and a dynamical system is a_1 = g(s_1), s_2 = h(a_1) and so on. Static systems are modelled by supervised/self-supervised learning, dynamical systems are modelled by reinforcement learning. LLMs at their core are trained in a self supervised way (the RLHF is afterwards, not on the core of the foundation model) so they cant model a situation where their actions can have consequences that they have to plan for. They dont model the idea of planning at all. Thats why they can write pretty well. And can even pretend to have furhter capabilities, but they fall short of actually being able to do things because that requires planning and understanding the dynamical nature of the real world Training an LLM or any modality of foundation model in a dynamical way would be the biggest improvement in AI since the attention mechanism